Aeon Atlas

A map of the aeon fork ecosystem — every public fork, what they enable, where they cluster.

This repo is a fork of aaronjmars/aeon whose sole purpose is to keep that map fresh. The weekly atlas refresh runs via the free no-LLM workflow .github/workflows/atlas-free.yml; the Claude atlas skill in aeon.yml stays disabled (empty Anthropic credits). Everything else from the upstream skill catalog ships disabled unless you enable it.

For what aeon itself is, see the upstream README.

Demo

Quartz graph of the fork ecosystem — three clusters, one per upstream root (aeon, aeon-agent, miroshark-aeon), connected by skill-overlap edges. Click the still for the 10-second physics animation (light version). Interactive version at swarm-ai-research.github.io/aeon-atlas/universe/. Re-record with node scripts/screenshot-universe.mjs --video --theme dark.

What you get

Artifact	What it shows
/universe/	Quartz-rendered graph + per-entity notes for every fork, ecosystem project, skill pack, and novel skill. Backlinks + tag clusters + global graph view.
docs/atlas.html	Interactive Cytoscape map. Fork tree, colored by recent activity. Click any node for its enabled-skill list, ★, last-push date.
docs/atlas.md	Readable digest — top forks by ★, most-active recent pushes, most-enabled skills across the fleet, strongest customization-overlap pairs.
docs/whats-new.md	Diff vs the most recent prior history/ snapshot — new/removed forks, ★ movers, new novel skills, ecosystem additions, activity transitions.
docs/digest/ + feed.xml	Permanent per-date weekly-digest pages and an Atom feed. Subscribe at /aeon-atlas/feed.xml.
atlas.json	Machine-readable: nodes (repos with metadata), edges (fork-of + skill-overlap), per-skill popularity.

How it builds the graph

For every public fork of aaronjmars/aeon (via the GitHub Forks API):

1. Fetch the fork's aeon.yml and parse which skills it has enabled: true.
2. Compute its delta from upstream (which skills it added or disabled vs. the upstream baseline). Vanilla forks — those that cloned and never customized — have an empty delta and don't get overlap edges. The signal of interest is customization patterns, not identical clones.
3. Score overlap between any two non-vanilla forks as the Jaccard similarity of their deltas. Keep the top 4 per fork above 0.30 with at least 2 shared customizations.

So the dense edges of the graph aren't "these forks both contain heartbeat" (universal — heartbeat ships enabled in upstream) but "these forks both enabled narrative-tracker and defi-monitor while disabling morning-brief."

Refresh cadence

Weekly, Sunday 04:00 UTC via Atlas (free) (.github/workflows/atlas-free.yml): node scripts/atlas.mjs regenerates everything, diffs against the prior run, and opens a PR only when something changed materially (new fork, ★ jump, dormant fork resumed, new high-overlap pair, etc.). No LLM keys required. The Claude atlas skill remains available in skills/atlas/SKILL.md if you enable it in aeon.yml with Anthropic credits.

Free paths for forks

Zero-key weekly atlas (no LLM) — this fork's Sunday path: .github/workflows/atlas-free.yml (Atlas (free), cron 0 4 * * 0 + workflow_dispatch). Enable GitHub Actions on your fork and keep that workflow. No Anthropic/Bankr/Ollama secrets required. Optionally add a repository secret DISCORD_WEBHOOK_URL to post when a PR opens; if unset, Discord is skipped quietly.

Free LLM skills (Claude Code via gateway — still needs a free-tier provider key):

1. In aeon.yml, set gateway.provider: openrouter or deepseek (default stays direct for paid Anthropic).
2. Add the matching Actions secret: OPENROUTER_API_KEY or DEEPSEEK_API_KEY.
3. Pin model: (and any per-skill model:) to a free/compatible id — Claude model names will not work on these routes.
   • OpenRouter examples: deepseek/deepseek-chat-v3-0324:free, meta-llama/llama-4-scout:free, openrouter/free
   • DeepSeek examples: deepseek-chat, deepseek-v4-flash, deepseek-v4-pro (pin from DeepSeek docs; IDs change)
4. Caveats: free OpenRouter models are rate-limited; tool/compatibility varies by model; Bankr/direct remain the paid paths.

To regenerate locally:

node scripts/atlas.mjs              # fresh fetch + write all artifacts
node scripts/atlas.mjs --cache      # reuse cached fork list if < 24h old
node scripts/atlas.mjs --json       # print atlas.json to stdout
node scripts/atlas.mjs --depth 2    # include forks-of-forks (default 1)

The script uses gh api for auth headroom; with no auth it falls back to unauthenticated curl (60 req/hr rate limit — usually fine for one run).

What it isn't

• Not a skill marketplace. It catalogs what's enabled in aeon.yml, not what's available in skills/. A fork can ship skills nobody runs; those don't show here.
• Not a contribution graph. Doesn't currently look at committers or contributor overlap — could be added later.
• Not real-time. Weekly cadence is enough for an ecosystem that adds a few forks per day.
• Not portable to other forks. The atlas always maps aaronjmars/aeon's descendants. Use --upstream owner/repo if you want to point it elsewhere.

Relationship to upstream

This is a vanilla GitHub fork — all upstream commits flow in via git fetch upstream main. The only divergence is scripts/atlas.mjs, skills/atlas/, the atlas line in aeon.yml, the rewritten README, and the regenerated artifacts. Periodic merges from upstream keep the rest in sync.

---

Originally created with the methodology pioneered in rsavitt/aeon's notegraph work (PRs #64–#79). Same TF-IDF + clustering tooling, applied to the fork network instead of internal notes.